Retrieval-based-Voice-Conversion-WebUI RTX 50XX support?

Nihao!

When to expect RTX 50XX support? 5070 TI and 5080 are not working at this moment.

Thank you!

Mar 07 '25 14:03 andivax

不兼容，我试了更新了torch都无法运行放弃吧！

Mar 08 '25 12:03 zxq3003

We need official answer from the developer.

Mar 08 '25 13:03 andivax

Despite the latest git update, RVC wasn't working for me, so I upgraded to torch-2.7.0.dev+cu128 to support my RTX 50XX GPU. While the GPU is now correctly recognized, the program fails with the following error: _pickle.UnpicklingError: Weights only load failed. It seems to be related to PyTorch 2.6+ changing torch.load() to default to weights_only=True, preventing RVC from loading the full model.

Is there a fix planned for this, or should we manually override this behavior in the code by setting torch.serialization.default_weights_only = False?

Mar 12 '25 19:03 omegabrain

Despite the latest git update, RVC wasn't working for me, so I upgraded to torch-2.7.0.dev+cu128 to support my RTX 50XX GPU. While the GPU is now correctly recognized, the program fails with the following error: _pickle.UnpicklingError: Weights only load failed. It seems to be related to PyTorch 2.6+ changing torch.load() to default to weights_only=True, preventing RVC from loading the full model.

Is there a fix planned for this, or should we manually override this behavior in the code by setting torch.serialization.default_weights_only = False? try wsl2，and you can refer to this https://github.com/RVC-Boss/GPT-SoVITS/issues/2026

Mar 14 '25 07:03 syd2333

has anyone had luck running this in wsl2 or docker?

Mar 16 '25 17:03 dirtmonster1337

I've changed the 314th line in checkpoint_utils.py into

with open(local_path, "rb") as f: state = torch.load(f, map_location=torch.device("cpu"), weights_only=False) # weights_only=False is for cuda 12.6+ It allows for loading checkpoints. Currently checking the training process with this fix.

Just to clarify, i did install pytorch for cuda 12.8.

Look like for training you also need to downgrade matplotlib. matplotlib==3.8.1 works for me so far.

Jun 12 '25 02:06 AdamRopelewski

will give this a shot

Jun 20 '25 12:06 dirtmonster1337

I've changed the 314th line in checkpoint_utils.py into

with open(local_path, "rb") as f: state = torch.load(f, map_location=torch.device("cpu"), weights_only=False) # weights_only=False is for cuda 12.6+ It allows for loading checkpoints. Currently checking the training process with this fix.

Just to clarify, i did install pytorch for cuda 12.8.

Look like for training you also need to downgrade matplotlib. matplotlib==3.8.1 works for me so far.

Thanks, I used this method to solve the problem.

If you don't want to modify the underlying fairseq code, you can also add a global registration before calling. Add this at line 20 of infer/lib/rtrvc.py: torch.serialization.add_safe_globals([fairseq.data.dictionary.Dictionary])

Jul 04 '25 17:07 Eachan35

how would you downgrade matplotlib on the built in?

Jul 04 '25 18:07 dirtmonster1337